Research map / Adversarial robustness and security / Certified and empirical robustness
Batchnorm Allows Unsupervised Radial Attacks
Amur Ghose, Apurv Gupta, Yaoliang Yu, Pascal Poupart
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| angular attack vs FGSM, deeper network and larger eps | Angular attack · ImageNet Resnet-152 eps=0.1 | top-1 accuracy: 9.82 -> 0.248 | Improved |
| angular attack vs FGSM on shallow net, small eps | Angular attack · ImageNet Resnet-18 eps=0.03 | top-1 accuracy: 1.95 -> 23.416 | Mixed |
| angular attack vs PGD (supervised) | Angular attack · ImageNet Resnet-18 eps=0.03 | top-1 accuracy: 0.006 -> 23.416 | Hurt |
| angular attack vs random noise | Angular attack · ImageNet Resnet-18 eps=0.03 | top-1 accuracy: 69.40 -> 23.416 | Improved |
| remove batchnorm (Fixup Resnet) instead of BN Resnet (ablation) | Angular attack · ImageNet Fixup Resnet | top-1 accuracy | Hurt |
| attack layer before BN layer instead of BN layer (ablation) | Angular attack · ImageNet Resnet-18 eps=0.03 | top-1 accuracy: 23.416 -> 27.3 | Hurt |
| grant access to deeper blocks (8 to 12) (ablation) | Angular attack · ImageNet Resnet-34 eps=0.03 | top-1 accuracy: 54.46 -> 6.40 | Improved |
| angular attack on ViT with LayerNorm vs FGSM | Angular attack · ImageNet ViT-B-16 eps=0.1 | top-1 accuracy: 35.1 -> 16.7 | Improved |
| angular attack vs best single-stage attack | Angular attack · ImageNet Resnet-18 eps=0.03 | top-1 accuracy: 30.45 -> 23.416 | Improved |
| vary initial iterations tinit 5-30 (ablation) | Angular attack · ImageNet Resnets eps=0.03 | top-1 accuracy | No clear effect |
| increase total iterations from 30 to 50 (ablation) | Angular attack · ImageNet Resnet-18 eps=0.03 | top-1 accuracy: -0.52 -> 0.38 | Improved |
| average angular loss over 4 layers instead of 1 (ablation) | Angular attack · ImageNet Resnet-18 eps=0.03 | top-1 accuracy: -1.37 -> 3.49 | Improved |
Closest papers by meaning
- Exploring Normalization in Deep Residual Networks with Concatenated Rectified Linear Units (AAAI 2017 · similarity 0.86)
- Deconstructing the Regularization of BatchNorm (ICLR 2021 · similarity 0.86)
- Backdoor Attacks Rerouted: BatchNorm as a Sink for Adversarial Signals (NeurIPS 2026 · similarity 0.85)
- Batch Normalization Orthogonalizes Representations in Deep Random Networks (NeurIPS 2021 · similarity 0.85)
- Strong Transferable Adversarial Attacks via Ensembled Asymptotically Normal Distribution Learning (CVPR 2024 · similarity 0.85)
- The Cost of Compression: Tight Quadratic Black-Box Attacks on Sketches for $\ell_2$ Norm Estimation (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
